Short-term spatio-temporal demand pattern predictions of trip demand

Being able to accurately predict future taxi demand can beneficial not only for taxi companies but also for passengers and the environment as as an intelligent taxi planning system can reduce waiting and idle driving times. This thesis proposes a simple LSTM based neural network structure for predicting future taxi demand for many locations at the same time. Experiments are performed on taxi data from New York City from June, 2016 to June, 2019 and also weather measurements from Central Park, NY for the same time. The model’s hyperparameters are tuned using a very simple selection method based on predictions for only one location at a time which results in a model that has very accurate results for most of the zones but fails to accurately predict the demand for zones with a very low average number of pickups. More complex, state-of-the-art algorithms, hyperopt and BOHB, implemented to tune the model’s hyperparameters in a more structured and comprehensive way can further reduce the error but only slightly (2.2% MSE reduction). The results suggest that there are factors that limit the performance gains of popular hyperparameter optimization techniques but also that a relatively simple model is able to yield useful predictions and outperform several
naive benchmark methods.